Intuit's Ashok Srivastava, on AI agents' new frontier

Ashok Srivastava, who leads Intuit's AI initiatives, predicted agentic AI's arrival two years ago. Now, he has his sights set on the future.

Srivastava was hired as Intuit's chief AI and data officer in 2017 to begin assembling its AI team and building what would become the company's generative AI operating system (GenOS) platform to support AI development. GenOS now includes GenStudio, which hosts both internally and externally developed large language models (LLMs); GenRuntime, which manages AI workloads, including agents; GenSRF, which is responsible for AI security, risk and fraud; and GenUX, a user interface.

The company behind TurboTax, Credit Karma, QuickBooks and Mailchimp built its own internal AI agent frameworks but is also an early contributor to recent community projects such as Google's Agent2Agent Protocol and Anthropic's Model Context Protocol.

In May, Intuit introduced a set of AI agents for QuickBooks, including Payments Agent, Accounting Agent, Finance Agent, Customer Agent and Product Management Agent. This week, it added support for those agents to its cloud-based Intuit Enterprise Suite.

Informa TechTarget spoke with Srivastava this month about the company's AI agent development approach and where he believes AI technology is headed next.

What was the biggest technical challenge to being an early mover in developing AI agents?

Ashok Srivastava

Ashok Srivastava: When we started on this, we knew that we would have to have GenRuntime work with multiple tools. The larger the number of tools, the harder it was for GenRuntime to do it. That's because it uses an LLM for reasoning, and the reasoning capability was -- and in many ways still is --nascent. Our APIs needed to be provisioned so that GenRuntime would know what they did, and we have thousands of APIs that we had to structure differently. If you have an API that's been defined by humans for human consumption, a lot of assumptions are made that a machine may not have, so the API needs to be made more robust.

Equally difficult is the challenge of changing people's mindsets to a world where semi-autonomous systems operate. Many developers have grown up building declarative code -- how do you change that paradigm to programming with prompts that have non-deterministic outputs from the LLM, from the standpoint of thinking about problems? We've created extensive training programs to make that happen.

There were some early cases with chatbots giving out inaccurate information. How do you ensure an AI agent doesn't go off the rails like that?

Srivastava: We developed GenSRF to evaluate model outputs based on criteria we choose. Part of it uses LLMs to assess the outputs of other LLMs. We also have dashboards and other tools so that our AI governance team can review the results. Our main goal is to create technologies that are safe and secure for our customers, which relies on a combination of technology, human evaluation and understanding. It also includes features to set guardrails for how the LLMs should behave.

In some cases, we've found that you can actually train the model to avoid [giving false or toxic answers]. But that can also be a great place to employ rules. We call those hybrid systems, and I'm a strong believer that hybrid systems are the way to go.

How do you decide which AI agent workflows should be deterministic, and which should be probabilistic? Are there any rules of thumb that have emerged there?

Srivastava: Personally, I think that there's absolutely a need for both rules-based systems and declarative code, as well as non-deterministic systems. I'm not a purist who says everything must be non-deterministic, nor do I say that everything should be deterministic. I think we should do the right thing for the right customer problem.

If you have a domain where a small set of rules can answer the need of the customer, just use it and move on. If you need reasoning across a very complex set of tasks, that's not a good place to use rules. For example, our Payments Agent helps businesses get paid five days faster than average. It can automatically look at the fact that you're creating an invoice, let's say for a customer that tends not to pay on time. To increase the likelihood that the customer pays you on time, the agent will suggest three options: One, increase the number of payment options; two, put in a fee if they pay late; three, make the tone of your message more assertive. That's a set of reasoning challenges that would be hard to approach through rules. The agent also ingests written data, pictures and other things to create the invoice that's sent out. Then it'll automatically say, 'I'm going to set up a reminder service so that after 10 days, if they haven't paid, they're going to get a reminder.' That capability leads to customers getting 10% more invoices paid in full.

What I see happening is increased automation and the building of a bridge between human and artificial intelligence.

You sensed AI agents coming a couple of years ago. What do you sense is coming for the mainstream next?

Srivastava: What I see happening is increased automation and the building of a bridge between human and artificial intelligence. We have a vast network of human experts assisting people with their finances, taxes, bookkeeping and more. There's a lot happening in this area, and it's an area that needs further exploration. How do we develop AI systems that can utilize human intelligence, and how do we create counterparts so humans can use AI to improve their work? This is a very important area for us to focus on moving forward.

Another key development is the emergence of better reasoning models. The question will be, how can they reason under uncertainty? When I know things for sure, reasoning becomes easier -- like if you're told there's going to be heavy rain outside, it's simple to decide to take an umbrella. But when faced with multiple options and ambiguous circumstances, human reasoning becomes much harder. It's even more challenging for machines. There is extensive research on this in AI, but not much on how to achieve it with large language models, which will be a critical area for future work.

Businesses must make constant decisions about their cash flow -- whether they'll have enough to pay bills, if a customer will pay on time or whether to take out a loan. The micro and macroeconomic factors are countless and uncertain. Humans respond differently to such uncertainty, but they can manage it. Now, we want machines to do the same and provide humans with insights to help them decide. That's the key link between human [intelligence] and AI -- creating algorithms that can operate with partial and often unknown information. This begins to bridge the gap between what are called Markov decision processes and intelligent agents.

Beth Pariseau, a senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism covering DevOps. Have a tip? Email her or reach out @PariseauTT.